Aiming at the non-linear and non-stationary characteristics of rolling bearing fault vibration signals, this paper proposes an adaptively bearing fault diagnosis method based on one-dimensional convolutional neural network (1DCNN) and long-short-term memory (LSTM) network. Firstly, the original one-dimensional vibration signals are input into the two channels of 1DCNN and LSTM by overlapping sampling. Then they fuse the feature information of spatial and temporal dimensions through the concatenate layer. Finally, the classifications of the fault categories are performed by the softmax classifier. The method can extract features from the original vibration signal adaptively and achieve “end-to-end” fault diagnosis. The experimental analysis of rolling bearings of different fault types, different sensor acquisition orientations and different fault diameters shows that the proposed method has higher recognition accuracy compared with other methods in identifying bearing fault categories, and has good adaptability and effectiveness.